Clastiq Orchestrator AI Platform

THE SOLUTION

One box, every team

Clastiq AI Platform is an NVIDIA GPU system with Clastiq Enterprise pre-installed. It gives each team a project of its own: members, roles, quotas, storage and keys. Teams deploy models, start databases and tools from the marketplace, run their own containers and notebooks, and fine-tune on the shared GPUs without ever touching the machine.

One vendor, one invoice, one support contract for hardware and software. The administrator decides how the GPUs are split. Every request, every GPU minute and every token is visible on one dashboard.

The model console: deployments, gateway traffic and GPU utilisation for the whole workstation on one screen, from one hour to seven days.

1/4 GPU

the smallest slice a notebook or embedding model can take, so eight cards serve dozens of jobs

4

inference engines: vLLM, SGLang, llama.cpp and TensorRT-LLM, chosen per deployment

1

OpenAI-compatible endpoint for every model, with keys and token quotas per project

0

bytes of weights, prompts or usage data leaving the box

GPU SHARING

Every GPU fully used, every team fairly served

The GPUs are the most valuable part of the box, and Clastiq treats them that way. The administrator gives each project a queue with a guaranteed GPU quota and a hard limit. When other teams are idle, a queue may borrow beyond its quota according to its weight. When demand returns, borrowed capacity is handed back.

• Fractional GPUs. A notebook, an embedding model or a small chat model can request a quarter or half of a card. Several workloads share one GPU safely, each with its own memory budget.
• Multi-GPU models. A 70B-class model spans four cards with tensor parallelism and still appears as one deployment.
• Queued, never refused. Work that exceeds the quota waits in the project's queue and starts the moment capacity frees up. Users see where their job stands.
• Autoscaling within limits. Deployments add replicas on load and give them back when it drops, never beyond the project's ceiling.

Whole cards for large models, fractions for small ones. Each team's queue has a guaranteed share, a hard ceiling and a weight for borrowing idle capacity.

OBSERVABILITY

See every GPU minute and every token

You paid for the GPUs; you should see them work. The model dashboard shows utilisation and memory per GPU and per deployment, request rates, token volume and latency, from the last hour to the last seven days. Project members see the same for their own workloads.

• Per deployment: GPU utilisation, GPU memory, replicas, requests per minute, tokens in and out, time to first token, errors.
• Per project: CPU, memory, GPU and storage against quota, with usage history and the workloads behind it.
• Per workload: live logs, events and resource graphs, and alerts when a model stops answering.
• For the administrator: the whole system at a glance, GPU temperatures and power included, plus ready-made monitoring dashboards for deeper analysis.

One deployment, one screen: GPU, memory, latency and logs side by side.

MODEL CONSOLE

Deploy any open model in minutes

From a Hugging Face checkpoint or a GGUF file to a production endpoint with autoscaling, guardrails and per-project keys. No infrastructure knowledge required from the team that needs the model.

MARKETPLACE

Everything around the model, one click away

A model alone is not an application. Teams need a vector database for retrieval, a relational database for state, a chat front end, an experiment tracker, a workflow engine. The marketplace starts each of them inside the project with a short form, on the same GPUs, storage and sign-on as the models.

• Vector databases: Qdrant, Milvus, Weaviate, Chroma.
• SQL and NoSQL: PostgreSQL, MariaDB, MongoDB, Redis, Elasticsearch, InfluxDB.
• AI applications: Open WebUI, Dify, Langflow, Langfuse, Ollama, Gradio, Streamlit.
• Data science and MLOps: JupyterLab, RStudio, MLflow, Aim, Optuna, Superset, Metabase, DataHub.
• Workflow and platform: Airflow, Dagster, NiFi, Argo Workflows, Keycloak, Authentik, container registries.

Every template is pinned to a tested version, validated on delivery, and updated with the platform. Administrators can add their own templates for in-house services.

Dozens of services, each started with a form: name, size, a password, done.

One project, one list: the model, the databases around it, your own code and the nightly job.

CUSTOM WORKLOADS

Bring your own containers

Anything that runs in a container runs on the box, next to the models and with the same GPU fractions, quotas and monitoring. Teams submit their own images without a ticket and without learning the infrastructure underneath.

• Services with replicas, ports and a public or internal address, such as a retrieval API in front of a vector database.
• Jobs and scheduled jobs for evaluations, data preparation, nightly re-indexing or batch inference, with a GPU slice when they need one.
• Virtual machines for software that is not containerised, with SSH access through the platform's sign-on.
• Three levels of control: a simple form (image, command, CPU, memory, GPU, ports, env), an advanced form (volumes, probes, autoscaling) or raw YAML for experts.
• Project resources attached: persistent volumes, object storage buckets, Hugging Face repos as folders, secrets and environment variables.

INFERENCE, GATEWAY & PLAYGROUND

From any open model to a production API

Pick a model from the catalog, pick the engine that runs it best, and it is live behind one OpenAI-compatible endpoint. Teams try it in the playground first, then copy the same request into their code with a project key.

• The right engine per model. vLLM, SGLang, llama.cpp or TensorRT-LLM, chosen per deployment with optimized presets: fractional GPUs for small models, tensor parallelism for large ones, fp8 KV cache and speculative drafts for throughput.
• One gateway for everything. Every model behind the same OpenAI-compatible API, so existing SDKs, agents and tools work unchanged. Project-scoped keys, tokens-per-minute quotas, routing across replicas and deployments.
• Playground before code. Chat with any gateway model using text, images, PDFs and code files. Compare answers, tune parameters, copy a working request.
• Usage you can see. Input and output tokens per project and per model, next to GPU time, so you always know which team is using what.

Same model, same key, same endpoint as the application that ships next week.

Playground example from the brief

supplier-contract-2026.pdf · 14 pages

Summarise the termination clauses and flag anything shorter than 90 days’ notice.

Two clauses allow termination. Clause 14.2 gives either party 180 days’ written notice. Clause 14.5 lets the supplier terminate on 30 days’ notice after a missed payment, which is below your 90-day threshold and should be renegotiated.

Ask a follow-up or attach a file…

AGENTIC AI ASSISTANT

An operator that diagnoses and acts, with your approval

Every console carries a chat panel that does more than answer questions. Ask why a workload is not running and it reads the status, restarts, last log lines, events and memory peaks, then returns findings with evidence and a fix. Ask it to deploy a model, scale an API or add a teammate, and it does it through the same permissions you have.

• Recognises the usual suspects. Out-of-memory kills, image pull errors, crash loops by exit code, pods no GPU can take, failed mounts, evictions, failing health checks, failed model downloads.
• Approval before any change. Each action or fix is a card with the exact call. Nothing runs until you approve, the approval is signed by the server, and every approved change is in the audit log.
• Never destructive, always in scope. It can create, start, stop, scale and change, never delete, and sees only what you could see yourself.
• Your model, on your box. Runs on a model served by the system itself or on any provider you choose. Voice input in the browser.
• Open to your agents. The same tools are exposed as an MCP endpoint, so coding agents and automation use the platform with a personal access token.

Assistant example from the brief

Assistant Project · product Why is api-gateway not running? I read its status, events, last log lines and memory history. Two findings: ! Out of memory. The container was killed (OOMKilled) 4 times in 20 minutes. Peak memory 3.9 GB against a 4 GB limit. ! Health check failing. Readiness on /healthz never passed because the process restarts first. ✓ Image pull, volumes and GPU scheduling are fine. The fix is a larger memory limit. 6 GB fits the project's quota.

Change workload update_workload workload: api-gateway ram: 4 GB → 6 GB · restart: yes [Approve] [Decline]

Approved. Also scale qwen3-32b to two replicas for the launch.

Scale deployment set_model_deployment_replicas deployment: qwen3-32b · replicas: 1 → 2 · GPUs: 2 → 4 (2 free in queue) [Approve] [Decline]

Ask or tell the assistant...

Findings with evidence, fixes and actions as approval cards: nothing runs until you say so.

BUILT FOR ONE ORGANIZATION

Your hardware. Your identity. Your data.

Everything runs on the box: models, prompts, datasets, fine-tuned weights, logs and metrics. Users sign in with your single sign-on, and projects keep teams' work apart.

Consoles & APIs

How people and automation use the platform

• Admin console
• Organization console
• Project console
• Model console
• REST API · OpenAPI
• MCP endpoint

IDENTITY AND QUOTAS ON EVERY CALL

Clastiq Enterprise

The model lifecycle, end to end

• Catalog & model store
• Deployments
• Gateway
• Fine-tuning
• Datasets
• Notebooks
• Guardrails
• Hugging Face volumes

SCHEDULED ONTO

NVIDIA hardware

Pre-installed, tuned and supported by Clastiq

• NVIDIA HGX or DGX GPUs with NVLink
• GPU scheduler with fractions & queues
• NVMe model & dataset storage
• Image registry
• Single sign-on
• Secrets
• Metrics & dashboards
• Backups

Projects for teams

Each team gets a project with its own members and roles, quotas for CPU, memory, GPU, storage and tokens, and its own API keys, buckets and secrets. Network policies keep projects apart on the shared machine.

Quotas you can trust

A project cannot take more than its limit, and never starves another team of its guaranteed share. Idle GPUs are lent out automatically and reclaimed when their owner needs them.

Operable by one admin

The box arrives installed and tuned for its GPUs. Upgrade in one step, back up and restore workloads from the console, and read every administrative change in the audit log. No platform team required.

AT A GLANCE

Feature list

NEXT STEP

Ready to run AI on your own terms?

A one-hour walkthrough covers the GPU class that fits your models, the teams that will share the box and how it connects to your sign-on and network. Your Clastiq AI Platform arrives installed and serves its first model on delivery day.

Book a walkthrough

Request the datasheet

Clastiq clastiq.ai

© 2026 Clastiq. NVIDIA, HGX, DGX, Grace, Hopper, Blackwell and ConnectX are trademarks of NVIDIA Corporation. Other names are trademarks of their respective owners.

vLLM, SGLang, llama.cpp, TensorRT-LLM and Hugging Face are independent projects. Specifications subject to change.